Skip to content

feat(windows): app icon + update service; fix taskbar freeze in text-selection flow - #101

Open
hubo1989 wants to merge 148 commits into
mainfrom
hubo/feat-windows
Open

hubo1989 wants to merge 148 commits into
mainfrom
hubo/feat-windows

Conversation

@hubo1989

@hubo1989 hubo1989 commented Nov 3, 2026

Copy link
Copy Markdown
Owner

What

Two commits:

1. e1875cc feat(windows): app icon, tray refinements, update service, MSIX packaging

  • App icon: App.rc + Assets (16–256px), wired into the window class and packaged manifest
  • Tray: icon restore on TaskbarCreated (explorer restart), CheckUpdates menu command via UpdateService
  • Packaging: build_msix.py (stage/sign/feed) + test
  • Build: /utf-8 on all vcxproj — fixes C4819→hard-error on zh-CN systems (code page 936); CI adjustments

2. 02ad4ff fix(windows): clipboard timeouts + shell-foreground guard prevent taskbar freeze

Root cause (reproduced on hardware): Ctrl+Alt+D/I text selection always takes the clipboard fallback (UIA is not wired yet). EmptyClipboard/GetClipboardData issue synchronous cross-process sends to the clipboard owner (WM_DESTROYCLIPBOARD / delayed-render WM_RENDERFORMAT) with no timeout. When the owner is explorer, the worker thread wedges forever and drags the shell UI thread (taskbar) down with it. Reproduced twice: worker thread stuck in Wait/Suspended with 0 CPU, taskbar unresponsive until the process is killed.

Fixes:

  • Win32TextExtractorHost: every clipboard transaction wrapped in RunClipboardOp — sacrificial detached thread, 800ms cap, abandoned thread parks in the system call instead of blocking
  • AppMain: text selection refused with a clear popup when the foreground process is a shell surface (explorer / start menu / search / input host / dwm) — Ctrl+C there copies files, never selectable text
  • AppMain: pipelineBusy watchdog re-arms after 15s so a dead worker can no longer lock out all hotkeys
  • Renamed Win32OcrUiaHosts.cpp → OcrUiaHostsWin32.cpp (was a scratch file name that snuck into the project)

Testing

  • Full MSVC Release x64 build green, Core tests pass
  • Signed MSIX (1.0.5.1) installed on the target Windows machine (zh-CN, self-signed CN=TransFrame Dev in TrustedPeople, no timestamp — RFC3161 server unreachable there)
  • Post-fix verification on that machine, tray_hung=False throughout: desktop Ctrl+Alt+D → refusal popup; 5× rapid-fire Ctrl+Alt+D; desktop Ctrl+Alt+I — app responsive, thread count stable, no shell freeze

Notes / remaining

  • No RFC3161 timestamp on the signature (network-blocked test box); renewal due 2027-09, distribution needs a real cert
  • UIA text-pattern extraction (focusedElement) is still unwired — text selection relies on the clipboard fallback; wiring it up will further reduce clipboard churn (follow-up)

hubo1989 and others added 30 commits November 2, 2026 02:20
…iefs

Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
…dow detect, hotkeys, tray (task B)

- CaptureService: WGC -> DXGI Desktop Duplication -> GDI BitBlt fallback
  chain, unified Core Screenshot output (physical px + DPI metadata),
  IsBorderRequired=false gated on UniversalApiContract 15 (Win10 yellow
  border documented degradation)
- DisplayMonitor: EnumDisplayMonitors + GetDpiForMonitor (Per-Monitor V2),
  2s cache refresh policy mirroring macOS DisplayFreshnessPolicy
- WindowDetector: EnumWindows + DWMWA_EXTENDED_FRAME_BOUNDS/CLOAKED with
  mouse-down snapshot cache (hover hit-test pure memory), FR-02 click vs
  4px drag threshold semantics
- SelectionOverlay: WS_POPUP+LAYERED+TOPMOST virtual-desktop window,
  D2D 1.1 + DComp rendering (40% dim, punch-through, 1px accent stroke,
  size badge, crosshair, window highlight), precreated for <=60ms hotkey
  path; portable MT-safe paint-list state machine
- HotkeyManager: RegisterHotKey/WM_HOTKEY x5 (Ctrl+Alt+F/A/T/D/I defaults),
  conflict reporting, runtime re-registration
- TrayIcon: Shell_NotifyIcon + menu model with recent-5 screenshot slots
- Catch2 v3.7.1 vendored; 60 test cases / 235 assertions green
  (geometry, DPI, cache policy, conflict matrix, fallback chain, overlay
  state machine, tray menu)
- PerMonitorV2 app.manifest; vcxproj (x64+ARM64, /W4 /WX) + CMake;
  per-service design docs in windows/docs/windows/

Windows build verification NOT RUN (macOS dev box); reproduce with
msbuild windows/TransFrame.sln or cmake -S windows/NativeShell.

Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Core (pure logic, platform-clean):
- 9 translation engines, no apple (PRD §6.12): MTranServer (probe ladder
  /health→/→/translate, 5s TTL), OpenAI/Claude/Gemini/Ollama via one
  LLMChatProvider (endpoint routing, HTTPS key guard, batching, truncation
  detection), Google v2 / DeepL (456 quota) / Baidu (MD5 sign, 54003/54004),
  Custom OpenAI-compatible with multi-instance UUID identity (custom:<uuid>
  credential slot, keyless localhost per macOS PR#92)
- TranslationScheduler: ordered chain ≤5 sequential fallback + aggregated
  errors, compare 2-4 parallel with per-run result cache, scene binding
  routing, cancellation aborts immediately
- EngineHealthChecker: four states, free probes only (never billed calls),
  60s TTL cache; Google/Baidu/Claude/Gemini/DeepL gate orange/gray on
  credentials without requests
- OCR: WindowsOcrEngine (language-pack gating, normalized bbox mapping),
  PaddleOcrEngine (fast/precise argv + JSON parsers, Python-literal
  normalization, health cache), VlmOcrEngine skeleton (5 provider slots,
  segment parsing)
- VisualLineGrouping port with macOS test vectors; prompt layering fixed
  to Swift order (scene > compatible > engine)
- TextExtractor (UIA GetSelection first, clipboard fallback with
  save/clear/Ctrl+C/poll/restore) + InsertionService (mandatory
  re-validation, never pastes on mismatch)

NativeShell (Win32, _WIN32-only, NOT RUN on this macOS host):
WinHttpTransport (WinHTTP 5.1, 10s timeout, cancel polling, error
taxonomy), CredManKeyStore (CredWrite/CredRead, TransFrame/<engine>/
<instance>), Win32OcrUiaHosts (WinRT OCR, PaddleOCR process runner,
UIA+SendInput hosts)

Tests: Catch2 v3, 105 cases / 370 assertions green (clang++ -std=c++20,
macOS arm64). Mock transports/providers drive request shapes, error-code
mapping, MD5 known vectors, scheduler matrices, health TTL, insertion
abort paths. Manual smoke entry tests/manual/smoke_translation.cpp with
MSVC build instructions.

Docs: docs/windows/engines/<engine>.md x9 (endpoints, params, error-code
tables) + windows/README.md build guide + tasks/briefs/report-C.md (FR
mapping, macOS-difference list).

Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Models: SettingsKeys (all AppSettings keys), KeyboardShortcut, all 11
Annotation types (tagged-JSON codec), Screenshot/CaptureDocument/DisplayInfo/
OCRResult/ScreenAnalysisResult (ImageRef handle for CGImage), BilingualSegment/
TranslationResult/EngineResult/TranslationResultBundle, TranslationHistory
(4 categories incl. color-picked convention), VLMProviderType/CompatibleConfig.

Algorithms: VisualLineGrouping.mergeLines/joinText, LLMTranslationBatching,
LLMTranslationResponseParser (OpenAI/Claude JSON+SSE, truncation detection),
TranslationPromptConfig placeholders - all ported from Swift with the same
test vectors (fixed a lineBox-union porting bug caught by the vectors).

Interfaces: IOcrEngine, ITranslationProvider, IHotkeyListener, IClipboard,
IHistoryStore, ICredentialStore (pure virtual, documented).

Solution: TransFrame.sln with Core/NativeShell(pch skeleton)/App(WinUI3
blank + MSIX manifest)/Tests(Catch2) projects, x64+ARM64 /W4 /WX.

Tests: 68 Catch2 cases / 261 assertions green on macOS clang++ 20
(includes fixture-file driven visual-line-grouping suite extracted from
TransFrameTests). CI: windows-ci.yml (msbuild dual-arch + ubuntu
portability check). Vendored dependency licenses/sources documented.

Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
… + 5th error category

Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
…badges

- Add configurable 'Screen Color Pick' global hotkey (default Ctrl+Cmd+C),
  wired through the full 5-hotkey convention: KeyboardShortcut default,
  AppSettings key/load/reset, HotkeyCoordinator registration, Settings
  recorder row with conflict detection, tray menu key equivalent
- Empty state now shows the live hotkey (rebind-aware) with a tray-path
  fallback (spec §4.2)
- Extract ColorPickFlow as the single completion pipeline for all three
  eyedropper surfaces (tray, preview toolbar, selection-in-place toolbar):
  hex normalize + history insert + toast with View History link (spec §4.5)
- HistoryStore: shared recordPickedColor/makeColorPickEntry entry; the
  .colorDidPick observer reuses it
- Swatch cell: hover actions and constant favorite star merged into one
  top-right cluster, no overlap (spec §4.2)
- Tests: +4 pipeline/hotkey cases; shortcut enum coverage extended to 6
  cases; full suite 325 passed / 0 failed; 10 strings lint OK

Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
hubo1989 and others added 22 commits November 9, 2026 00:52
四通道请求形状自检(签名/URL/头/multipart);本地 python HTTP 替身承载
PicGo 全链路(真实 WinHttpTransport + 临时文件写入/清理 + 进程回收);
取色器程序化驱动(开窗→模拟移动/点击→与隐藏后直采像素对拍→Esc 取
消);取色历史临时库驱动(网格行数/搜索/收藏/删除/清空仅动 color 类
别)。44 项断言全 PASS;配套 build_smoke_uploads_picker.bat。

Co-Authored-By: Claude Code <noreply@anthropic.com>
… smoke

- LocalizationTests: codec roundtrip (CJK + surrogate pair), %@ format
  marker, table parse accept/reject, fallback chain (selected -> en ->
  default), empty-table clearing, switch listeners, system-locale
  provider mapping, loadDirectory contract, and the shipped-table schema
  (10 languages, identical key sets, non-empty values; tables copy to
  the test output dir via vcxproj Content)
- OnboardingTests: choice info mapping, privacy classification
  (localhost/mtran/ollama = local; disabled never sends), default-chain
  merge preserving unrelated keys, done-flag round-trip, hotkey defaults
  + per-action overrides, 3-step flow machine walk
- manual smoke_locale_onboard: L1/L2 table contract (259 keys, spot
  values zh/en/ja), L3 fallback + listener, O1 system locale probe,
  O2 onboarding flow + engines.json merge + done flag, N1/N2
  notification seam + live tray balloon (N/A when no tray session);
  build_smoke_locale_onboard.bat wraps the single-line cl invocation

Core suite 6437 assertions / 193 cases green (baseline 1057/172);
NativeShell 455/93 (baseline 445/92); all four manual smokes PASS.

Co-Authored-By: Claude Code <noreply@anthropic.com>
# Conflicts:
#	windows/App/AppMain.cpp
#	windows/Core/TransFrame.Core.vcxproj
#	windows/NativeShell/TransFrame.NativeShell.Tests.vcxproj
#	windows/NativeShell/src/PreviewWindow.cpp
#	windows/NativeShell/src/TrayMenu.cpp
#	windows/NativeShell/tests/TrayMenuTests.cpp
Leftover conflict marker at the end of the localization TEST_CASE
(caught by /WX build); merged menu now has 13 items - pickColor and
color-history joined the localized labels with tw() fallbacks.

Co-Authored-By: Claude Code <noreply@anthropic.com>
- 16 new keys across all 10 string tables (tray pickColor/colorHistory,
  preview upload, picker hint, color-history title/empty/copied/
  clear-confirm/context-menu/toggles/count format); tables re-serialized
  with the original 1-space indent so the diff is exactly the added keys
- ColorPicker/ColorHistoryWindow swap hardcoded zh labels for tw() with
  zh fallbacks (runtime identical in zh, localized elsewhere)
- fix a real crash found by the smoke: D2D1 DrawTextW does not take the
  GDI -1 length convention - static_cast<UINT32>(-1) reads past the
  buffer (0xC0000005 at first render); pass hint.size() instead
- smoke_uploads_picker: unbuffered stdout so a mid-run crash no longer
  swallows the stage log

Co-Authored-By: Claude Code <noreply@anthropic.com>
The minimal tw() repro's object file slipped into a65b2d3's 'add -A';
untrack it and ignore *.obj at the repo root (sources of truth are all
under tracked dirs; build outputs already live under ignored build/).

Co-Authored-By: Claude Code <noreply@anthropic.com>
…gning

- build_msix.py: --self-signed-test-cert (pwsh-first, falls back to
  powershell.exe; creates a disposable cert in CurrentUser\My and trusts
  it in CurrentUser\Root so signtool verify /pa passes on the machine);
  --certificate-thumbprint now mutually exclusive with it; stage-content
  self-check (EXE non-empty PE, per-arch bootstrap DLL, declared logos,
  Assets\Strings\*.json, VCLibs manifest dependency) before makeappx
  so packaging fails loudly instead of shipping a broken package.
- test_build_msix.py: StageContentTests cover missing bootstrap DLL,
  missing localization tables, missing declared logos.
- README: document the test-cert flag as local/CI-only.

Co-Authored-By: Claude Code <noreply@anthropic.com>
Per-arch leg: sln Release build -> VCLibs Desktop appx (fixed version,
fail-fast on download errors) -> build_msix.py --self-signed-test-cert
-> upload-artifact named msix-x64 / msix-arm64 (msix + .appinstaller +
vclibs appx). Test-signed artifacts are install-test only.

Co-Authored-By: Claude Code <noreply@anthropic.com>
- SettingsShell: gate Mica SystemBackdrop behind supportsMicaBackdrop()
  (OS build >= 22000) with try/catch; Win10 branch explicitly sets the
  ElementTheme::Default solid fallback so the window chrome does not
  render blank on the no-op backdrop.
- wgcSupportsBorderless(): belt-and-braces — the OS build probe must
  agree with the API-contract version, so a contract shim on Win10 can
  never claim borderless capture (yellow border stays as documented).
- New tf/PlatformFeatures.hpp + .cpp: testable version/build policy
  probes; Catch2 cases cover 19044 / 21999 / 22000 / 22621 boundaries.
- docs/windows/win10-degradation-matrix.md: static audit matrix (API ->
  Win10 behavior -> guard location -> verification method); remaining
  risk note for AppNotificationManager packaged backend TODO.

Co-Authored-By: Claude Code <noreply@anthropic.com>
…tion

SmartRedactionEngine (Core, pure): macOS SmartPrivacyRedactionService /
performSmartRedaction port. Privacy prompt pair aligned with
VLMProvider.swift VLMPromptPair.privacy; parser handles chat-completion
content, json fences, segments/results envelopes, truncation salvage
(quote/escape aware), NaN drop, clamp + degenerate drop, configurable
minConfidence (0.5 macOS default, missing confidence drops), pixel-space
output with padPx edge clipping and CGRect.integral rounding.

AnnotationDocument.addAll() appends a batch as ONE undo transaction so a
single Ctrl+Z clears the whole mosaic batch (empty batch: no history entry).

credentialAccount() uses the EnginesPage schema (TransFrame/vlm/<p>);
isConfigured() also accepts the legacy vlm_api_key_<p> account
VlmOcrEngine reads, so either writer enables FR-10.

Co-Authored-By: Claude Code <noreply@anthropic.com>
…t gate

SmartRedactionTests: prompt-pair anchors, Configuration gating (both
credential schemes), content/network error passthrough, pixel mapping,
fence unwrap, truncation salvage incl. braces-inside-strings, clamp/de-
generate drop, NaN + low-confidence filter, pad clipping, batch-as-one-
undo-step with redo, empty batch pushing no history.

CrossPlatformSnapshotTests: 5 fixture-driven TEST_CASEs with FULL-field
equality (bbox edges 1e-9, confidence 1e-6, text, order — never sampled):
- visual-line-grouping.json: 12 merge cases + 5 joinText cases
- llm-response-truncation.json: 3 batching + 4 parser cases
- translation-configuration.json: 2 prompt resolution cases
- ordered-chain-translation.json: 5 chain + 2 resolvedEngine cases
Vectors extracted from macOS TransFrameTests sources (each file's
_source/_macOSWiringNote names the Swift origin + future loader point).

Registrations: Core.Tests.vcxproj (ClCompile + Content fixture copy —
Content+CopyToOutputDirectory actually copies; the old None metadata
never did, fixed for the pre-existing fixture entry too) and CMake
TEST_SRC + POST_BUILD fixture copy. CMake CORE_MODEL_SRC now mirrors
the vcxproj Core-root sources (tf::core VisualLineGrouping et al. never
linked there before — pre-existing gap surfaced by the new tests).

Co-Authored-By: Claude Code <noreply@anthropic.com>
…ings

PreviewWindow: SmartRedact toolbar button (ToolButton 112) with
setSmartRedactHandler (base pixels, AppMain owns the worker) and
applySmartRedaction(batch) — document.addAll + history sync + repaint,
one Ctrl+Z clears the whole batch. PreviewWindow.cpp diff: +16 lines.

App: SmartRedactionFlow.{hpp,cpp} owns the worker body (OpenAI-
compatible chat.completions vision call covering the OpenAI slot and
OpenAI-mode endpoints incl. Ollama /v1; Claude/Gemini shapes land with
FR-17) so AppMain keeps only the wiring — one kSmartRedactDone case +
one setSmartRedactHandler callback (+62 lines). Unconfigured VLM
surfaces a Configuration notice and opens the engine settings page;
empty detection is an Info notice; batch applies as one undo step.

Strings: win.preview.smartRedact{,.done,.none,.failed,.noImage} added
to all 10 shipped tables (schema test enforces identical key sets).

Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Code <noreply@anthropic.com>
signtool verify /pa requires a chain to a trusted root, so it always
fails for the self-signed test certificate on local sideload flows where
the cert is trusted in TrustedPeople instead (sufficient for
Add-AppxPackage, invisible to /pa). The run then ended in a
CalledProcessError after producing a perfectly valid package.

Downgrade the final /pa failure to an explanatory note on the
--self-signed-test-cert path only; CA-signed production packages keep
the strict check. CI (elevated, LocalMachine/Root trust) still passes
/pa unchanged.

Co-Authored-By: Claude Code <noreply@anthropic.com>
…audit harness

Screenshot-driven UI audit at the macOS-parity 800x600 default found every
tab clipping its right edge: pages laid out for a ~720px column while the
WinUI NavigationView default pane (320px) left only ~430px of viewport.

- SettingsUiKit MakePageRoot: stretch the page column and disable
  horizontal scrolling so content wraps/shrinks to the real viewport.
- SettingsShell: OpenPaneLength 232 (macOS sidebar is ~200pt) restores a
  ~520px viewport; add SwitchSettingsTab test hook (same dispatch pattern
  as RunSettingsTabsSmoke, one tab per call, caller owns pacing).
- LanguagePage: drop the duplicated interface-language card; macOS keeps
  AppLanguagePicker only in GeneralSettingsTab, so parity says the same.
- AdvancedPage: show the expanded real logs path instead of the raw
  %LOCALAPPDATA% literal (matches the data-dir row).
- Tests/manual: smoke_settings_shots harness - per-tab switching with
  localization loaded like AppMain, offscreen PrintWindow capture
  (PW_RENDERFULLCONTENT, blank-check + BitBlt fallback), PNG via WIC.

Verified: six tabs re-shot at 800x600 with no right-edge clipping; tabs
smoke still 6/6.

Co-Authored-By: Claude Code <noreply@anthropic.com>
Per-page screenshot audit against the SwiftUI settings surfaced chrome
and material gaps:

- Custom title bar: ExtendsContentIntoTitleBar with a centered tab-name
  strip (mac navigationTitle) and theme-aware caption button colors; the
  native light title bar no longer clashes with the dark theme.
- Fixed 200px sidebar with Segoe Fluent icons per tab (SF Symbol
  mapping from SettingsTab.swift), no hamburger, PaneClosing cancelled
  (mac NavigationSplitViewColumnWidth min=ideal=max=200).
- Window is fixed-size like the mac styleMask (no .resizable);
  maximize disabled (1.6 lacks HasMaximizeButton, so it shows disabled).
- Cards follow DesignSystem.macosGroupedCard: radius 12, padding 16,
  controlBackground fill, hairline separator stroke.
- Page column spans the viewport (mac maxWidth: .infinity), and the
  in-page duplicate tab headers moved into the title bar.
- Kit adds MakeSettingRow (LabeledContent-style label-left /
  control-right rows) and MakeDivider for the upcoming per-page
  material pass.
- RunSettingsTabsSmoke / SwitchSettingsTab use the nav reference
  directly; window content is now a Grid (title strip + nav).

Verified: 6/6 tabs smoke; six screenshots re-audited clean at 800x600
(icons, centered per-tab title, no collapse gap, expanded logs path).

Co-Authored-By: Claude Code <noreply@anthropic.com>
Second material pass from the per-page screenshot audit. The SwiftUI
settings use LabeledContent-style rows (label left, control right,
hairline dividers); the port stacked labels above controls.

- GeneralPage: default format, export quality, interface language, and
  the startup/update toggles become MakeSettingRow rows with a divider;
  toggles drop the WinUI default On/Off captions (mac switches are bare)
  and carry hint lines in the row label instead.
- AdvancedPage: stroke width and text size sliders move into rows with
  right-aligned values, divider between.
- Strings: two new row hints (win.settings.autostart.hint,
  win.settings.checkUpdates.hint) across all 10 language tables; the
  rewrite also re-sorted a few misordered existing keys.

Engines/Prompts/Language/Shortcuts keep their current shapes: the macOS
tabs themselves stack labels above fields there (forms and rows with
trailing buttons already match).

Verified: 6/6 tabs smoke; screenshots re-audited at 800x600.

Co-Authored-By: Claude Code <noreply@anthropic.com>
Comment on lines +36 to +41
- uses: actions/checkout@v4

- name: Add MSBuild to PATH
uses: microsoft/setup-msbuild@v2

- name: Build solution

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

performance · medium
所有 job 均未配置任何缓存(NuGet restore、CMake 编译产物、apt 安装的 g++-13),每次运行都全量 restore + 双平台(x64/ARM64)编译,CI 时长与 runner 资源消耗偏高。建议对 msbuild/restore 使用 actions/cache(缓存 NuGet 包目录),对 CMake 构建目录和 ccache 做缓存。

Suggestion:

Suggested change
- uses: actions/checkout@v4
- name: Add MSBuild to PATH
uses: microsoft/setup-msbuild@v2
- name: Build solution
- uses: actions/checkout@v4
- name: Cache NuGet packages
uses: actions/cache@v4
with:
path: ~/.nuget/packages
key: nuget-${{ runner.os }}-${{ hashFiles('windows/**/*.proj', 'windows/**/*.sln') }}
- name: Add MSBuild to PATH
uses: microsoft/setup-msbuild@v2
- name: Build solution

Comment on lines +38 to +39
- name: Add MSBuild to PATH
uses: microsoft/setup-msbuild@v2

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

security · medium
microsoft/setup-msbuild 是第三方(非 actions/* 官方组织)维护的 action,仅按可变 tag v2 引用存在被劫持风险。建议固定到完整 commit SHA(与 checkout/upload-artifact 一致处理,或至少对该第三方 action 固定 SHA)。

Suggestion:

Suggested change
- name: Add MSBuild to PATH
uses: microsoft/setup-msbuild@v2
# 建议固定为具体 commit SHA(示例,请替换为最新验证过的 SHA)
- name: Add MSBuild to PATH
uses: microsoft/setup-msbuild@<full-commit-sha>

Comment on lines +163 to +164
-o ../build-tests 2>&1 | head -400
../build-tests

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · medium
编译输出通过 2>&1 | head -400 截断。当编译错误超过 400 行时,g++ 会因管道关闭收到 SIGPIPE 而被终止:真实编译错误被静默丢弃,且 set -o pipefail 下步骤会以 SIGPIPE(141) 失败而非真实的编译错误码,日志中也看不到根因。建议将输出写入临时文件再截断展示,保证退出码与完整错误日志都保留。

Suggestion:

Suggested change
-o ../build-tests 2>&1 | head -400
../build-tests
log=$(mktemp)
if ! g++-13 -std=c++20 -Wall -Wextra -Werror \
-I . -I Core -I external \
Core/*.cpp external/catch2/catch_amalgamated.cpp \
Tests/TestMain.cpp Tests/CoreModelsTests.cpp Tests/LLMTranslationCoreTestsA.cpp \
Tests/TranslationPromptConfigTests.cpp Tests/VisualLineGroupingTests.cpp \
-o ../build-tests > "$log" 2>&1; then
tail -400 "$log"; rm -f "$log"; exit 1
fi
rm -f "$log"
../build-tests

Tests/TestMain.cpp Tests/CoreModelsTests.cpp Tests/LLMTranslationCoreTestsA.cpp \
Tests/TranslationPromptConfigTests.cpp Tests/VisualLineGroupingTests.cpp \
-o ../build-tests 2>&1 | head -400
../build-tests

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · low
portable-check 任务把测试可执行文件输出到仓库根目录(../build-tests),且两个 CMake 构建目录分别落在 windows/NativeShell/build 与 windows/build-c。根 .gitignore 的 build/ 模式能匹配 windows/build 和 windows/NativeShell/build,但既不匹配根目录的 build-tests 文件,也不匹配 windows/build-c 目录。CI/本地构建产物会以未跟踪文件形式污染 git status,容易被误提交。建议将构建输出统一放入已被忽略的目录(如 windows/build/ 下),或在 .gitignore 中补充 build-tests 与 build-c 条目。

Suggestion:

Suggested change
../build-tests
../windows/build/build-tests

$zip = [System.IO.Compression.ZipFile]::OpenRead((Resolve-Path $out))
try {
$entry = $zip.GetEntry('AppxManifest.xml')
if (-not $entry) { Write-Error "$out has no AppxManifest.xml"; exit 1 }

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

style · low
脚本中多处使用 Write-Error ...; exit 1 的混合错误处理。GitHub Actions 的 pwsh 默认 $ErrorActionPreference = 'Stop',Write-Error 会先抛出 terminating error 使 exit 1 不可达,两种失败路径行为不一致、可读性差。建议统一使用 throw(与上方步骤风格一致)。

Suggestion:

Suggested change
if (-not $entry) { Write-Error "$out has no AppxManifest.xml"; exit 1 }
if (-not $entry) { throw "$out has no AppxManifest.xml" }

Comment on lines +177 to +179
// Placeholder until NativeShell exposes image info; dims flow through
// CaptureDocument::baseImage instead.
return Size{0, 0};

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · medium
Screenshot::dimensions() 在 image != 0 时仍恒返回 Size{0, 0}(占位实现),导致 estimatedFileSize / formattedFileSize / aspectRatio 全部返回 0 或 "0 B"。注释说明依赖任务 B 接线,但在接线前所有调用方会得到错误的文件大小估算;建议加显式 TODO 并在上层避免使用这些结果,或直接断言未实现以暴露误用。

Suggestion:

Suggested change
// Placeholder until NativeShell exposes image info; dims flow through
// CaptureDocument::baseImage instead.
return Size{0, 0};
// TODO(task B): wire real ImageRef -> dimension lookup via NativeShell.
// Interim: always 0x0 so callers must not treat these values as real.
return Size{0, 0};

Comment on lines +70 to +71
if (kind.contains("builtin")) {
const std::string raw = kind.at("builtin").at("_0").get<std::string>();

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · medium
fromJson 在 kind 为 {"builtin": null} 或缺少 "_0" 键时,kind.at("builtin").at("_0") 会抛出底层库异常,异常信息缺乏上下文且调用方难以区分是配置损坏还是代码缺陷。建议对两个分支的 "_0" 访问做存在性检查并抛出带说明的 std::invalid_argument,与下方 builtin 引擎名校验风格保持一致。

Suggestion:

Suggested change
if (kind.contains("builtin")) {
const std::string raw = kind.at("builtin").at("_0").get<std::string>();
if (kind.contains("builtin")) {
if (!kind.at("builtin").contains("_0")) {
throw std::invalid_argument("EngineChainItem builtin kind missing payload");
}
const std::string raw = kind.at("builtin").at("_0").get<std::string>();

Comment on lines +51 to +53
static KeyboardShortcut translateAndInsertDefault() {
return {0x22, CarbonModifiers::cmdKey | CarbonModifiers::shiftKey};
}

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · medium
C++ 侧缺少与 macOS 对齐的取色快捷键:Swift 的 KeyboardShortcut.pickColorDefault 与 AppSettings.Keys.pickColorShortcut 已存在(本次 macOS 变更中新增),但这里没有对应的 default 工厂方法,SettingsKeys.h 也没有 pickColorShortcut 键。从 macOS 导出的设置在 Windows 还原时该项会被静默丢弃,且后续 Windows 取色热键功能(B 任务)接入时需要回补此模型。建议同步补充。

Suggestion:

Suggested change
static KeyboardShortcut translateAndInsertDefault() {
return {0x22, CarbonModifiers::cmdKey | CarbonModifiers::shiftKey};
}
static KeyboardShortcut translateAndInsertDefault() {
return {0x22, CarbonModifiers::cmdKey | CarbonModifiers::shiftKey};
}
static KeyboardShortcut pickColorDefault() {
return {0x2E /* c */, CarbonModifiers::cmdKey | CarbonModifiers::shiftKey};
}

Comment on lines +71 to +76
static KeyboardShortcut fromJson(const Json& j) {
KeyboardShortcut s;
s.keyCode = static_cast<std::uint32_t>(j.at("keyCode").get<std::uint64_t>());
s.modifiers = static_cast<std::uint32_t>(j.at("modifiers").get<std::uint64_t>());
return s;
}

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · medium
fromJson 对所有字段使用 j.at(key).get(),在键缺失或类型不匹配时抛出 nlohmann 异常。该函数用于解析持久化设置(可能来自 macOS 导出、旧版本或手改文件),若调用方未捕获,单个坏字段会导致设置加载整体失败甚至启动崩溃。文件内已有 getOptional/getOptionalOr 辅助,建议对可回退字段使用缺省值,或在调用点统一 try/catch 并回退到默认快捷键。

Suggestion:

Suggested change
static KeyboardShortcut fromJson(const Json& j) {
KeyboardShortcut s;
s.keyCode = static_cast<std::uint32_t>(j.at("keyCode").get<std::uint64_t>());
s.modifiers = static_cast<std::uint32_t>(j.at("modifiers").get<std::uint64_t>());
return s;
}
static KeyboardShortcut fromJson(const Json& j) {
KeyboardShortcut s;
s.keyCode = static_cast<std::uint32_t>(getOptionalOr<std::uint64_t>(j, "keyCode", 0));
s.modifiers = static_cast<std::uint32_t>(getOptionalOr<std::uint64_t>(j, "modifiers", 0));
return s;
}

Comment thread windows/Core/Styles.h
Comment on lines +34 to +37
static CodableColor fromJson(const Json& j) {
return {j.at("red").get<double>(), j.at("green").get<double>(),
j.at("blue").get<double>(), j.at("opacity").get<double>()};
}

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · medium
fromJson 同样依赖 j.at(...).get(),styles 数据(如 shapeFillColor、strokeColor)损坏时会抛异常;这些默认值本可安全回退(红色/白色等),建议用 getOptionalOr 兜底,避免设置加载因局部损坏而整体失败。

Suggestion:

Suggested change
static CodableColor fromJson(const Json& j) {
return {j.at("red").get<double>(), j.at("green").get<double>(),
j.at("blue").get<double>(), j.at("opacity").get<double>()};
}
static CodableColor fromJson(const Json& j) {
return {getOptionalOr<double>(j, "red", 0.0), getOptionalOr<double>(j, "green", 0.0),
getOptionalOr<double>(j, "blue", 0.0), getOptionalOr<double>(j, "opacity", 1.0)};
}

…+ key hints

The selection overlay self-dismissed on WM_KILLFOCUS: under the Win32
foreground lock the SetFocus/SetForegroundWindow grab right after
ShowWindow bounces and the focus loss killed every session within
milliseconds - hotkeys and the tray menu's capture items ran but looked
completely dead. Remove the focus-loss dismissal (mac nonactivating
NSPanel parity; Esc / click-away / explicit cancel remain) and grab
keyboard focus via AttachThreadInput so Esc reliably reaches the
overlay.

Tray menu styling on dark systems: opt the HMENU into the Win11 dark
context menu via SetPreferredAppMode following AppsUseLightTheme, and
render each capture row's live key equivalent right-aligned
("全屏截图\tCtrl+Alt+F") like the mac status menu; hints refresh after
applySpecs and hotkey settings saves.

Verified live: 5/5 hotkeys registered, 3/3 injected region sessions
stay visible past the 2s focus-bounce window and dismiss on Esc.

Co-Authored-By: Claude Code <noreply@anthropic.com>
Comment on lines +2 to +3
<script type="text/x-dc" data-dc-script data-props='{"$preview":{"width":880,"height":560}}'>class Component extends DCLogic {}</script>
</body></html> No newline at end of file

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

style · low
文件末尾缺少换行符(EOF newline)。建议在文件结尾补上换行,避免在版本控制与文本工具中产生 "\ No newline at end of file" 提示,也与常见代码规范保持一致。

Suggestion:

Suggested change
<script type="text/x-dc" data-dc-script data-props='{"$preview":{"width":880,"height":560}}'>class Component extends DCLogic {}</script>
</body></html>
+<script type="text/x-dc" data-dc-script data-props='{"$preview":{"width":880,"height":560}}'>class Component extends DCLogic {}</script>
+</body></html>

Comment on lines +41 to +46
- name: Build solution
# -restore: the solution references WASDK/CppWinRT via
# PackageReference (WinUI3Shell + App); a fresh runner has no
# packages until NuGet restore runs.
working-directory: windows
run: msbuild TransFrame.sln /m /restore /p:Configuration=${{ matrix.configuration }} /p:Platform=${{ matrix.platform }} /warnaserror /verbosity:minimal

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

performance · medium
三个 Windows job 每次 /restore 全量拉取 WASDK/CppWinRT NuGet 包,CMake leg 也无任何构建目录缓存,windows-latest 上这类大型依赖还原显著拖慢 CI。建议为 NuGet 引入 actions/cache(缓存 %LOCALAPPDATA%\NuGet\packages,key 基于 windows//.csproj/**.vcxproj 哈希)或使用 setup-msbuild 之后配合 NuGet cache;CMake 构建目录缓存收益有限,可不做。

Suggestion:

Suggested change
- name: Build solution
# -restore: the solution references WASDK/CppWinRT via
# PackageReference (WinUI3Shell + App); a fresh runner has no
# packages until NuGet restore runs.
working-directory: windows
run: msbuild TransFrame.sln /m /restore /p:Configuration=${{ matrix.configuration }} /p:Platform=${{ matrix.platform }} /warnaserror /verbosity:minimal
- name: Cache NuGet packages
uses: actions/cache@v4
with:
path: ~\AppData\Local\NuGet\packages
key: nuget-${{ matrix.platform }}-${{ hashFiles('windows/**/*.proj', 'windows/**/*.vcxproj') }}
restore-keys: nuget-${{ matrix.platform }}-
- name: Build solution
working-directory: windows
run: msbuild TransFrame.sln /m /restore /p:Configuration=${{ matrix.configuration }} /p:Platform=${{ matrix.platform }} /warnaserror /verbosity:minimal

Comment on lines +158 to +163
g++-13 -std=c++20 -Wall -Wextra -Werror \
-I . -I Core -I external \
Core/*.cpp external/catch2/catch_amalgamated.cpp \
Tests/TestMain.cpp Tests/CoreModelsTests.cpp Tests/LLMTranslationCoreTestsA.cpp \
Tests/TranslationPromptConfigTests.cpp Tests/VisualLineGroupingTests.cpp \
-o ../build-tests 2>&1 | head -400

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · high
portable-check 的 "Core (task A)" 步骤用内联 g++ 命令只编译了 5 个测试 TU,实际只覆盖极少量 Core 源文件;而 windows/CMakeLists.txt 中 translation_tests 通过 file(GLOB Core/src/*.cpp) + CORE_MODEL_SRC 已包含全部可移植源(HistoryStore.cpp、EngineConfigStore.cpp、OcrEngines.cpp、SmartRedaction.cpp 等均在 CMake leg 中构建)。结果是:-Werror(无 Win32 泄漏的强保证)只作用于这一小撮文件,而完整源集只在无 -Werror 的 CMake leg 中编译——可移植性检查的强度与覆盖面错位。此外 2>&1 | head -400 在 pipefail 下,编译器因 head 提前退出收到 SIGPIPE 被杀时退出码语义混乱,可能掩盖真实编译错误或产生误导性失败。建议:要么让 task A leg 直接复用 CMake 并加 -Werror(如 -DCMAKE_CXX_FLAGS=-Werror),要么把完整源清单写入脚本并移除 head 截断(改用输出重定向到文件再 tail)。

Suggestion:

Suggested change
g++-13 -std=c++20 -Wall -Wextra -Werror \
-I . -I Core -I external \
Core/*.cpp external/catch2/catch_amalgamated.cpp \
Tests/TestMain.cpp Tests/CoreModelsTests.cpp Tests/LLMTranslationCoreTestsA.cpp \
Tests/TranslationPromptConfigTests.cpp Tests/VisualLineGroupingTests.cpp \
-o ../build-tests 2>&1 | head -400
# 让 GCC leg 覆盖与 CMake 相同的完整可移植源集,并保留 -Werror
# (源清单维护在 windows/CMakeLists.txt 或独立脚本中,避免两处漂移)
g++-13 -std=c++20 -Wall -Wextra -Werror ... -o ../build-tests
../build-tests

Comment on lines +158 to +163
g++-13 -std=c++20 -Wall -Wextra -Werror \
-I . -I Core -I external \
Core/*.cpp external/catch2/catch_amalgamated.cpp \
Tests/TestMain.cpp Tests/CoreModelsTests.cpp Tests/LLMTranslationCoreTestsA.cpp \
Tests/TranslationPromptConfigTests.cpp Tests/VisualLineGroupingTests.cpp \
-o ../build-tests 2>&1 | head -400

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · low
portable-check 的 g++ 步骤用 2>&1 | head -400 截断输出:当编译错误超过 400 行时,位于前面的关键报错(MSVC/GCC 通常第一条错误在最前,尚可保留,但后续模板展开类错误会被截断),更重要的是 head 提前退出会触发 SIGPIPE,配合上面 set -euo pipefail 时,即使 g++ 正常完成、仅因输出量大被 head 掐断管道,整个步骤也可能以非零退出误报失败。建议去掉 head 截断(CI 日志本身会完整留存),或改用 g++ ... || { tail -n 400 g++.log; exit 1; } 这类先落盘再截取的方式。

Suggestion:

Suggested change
g++-13 -std=c++20 -Wall -Wextra -Werror \
-I . -I Core -I external \
Core/*.cpp external/catch2/catch_amalgamated.cpp \
Tests/TestMain.cpp Tests/CoreModelsTests.cpp Tests/LLMTranslationCoreTestsA.cpp \
Tests/TranslationPromptConfigTests.cpp Tests/VisualLineGroupingTests.cpp \
-o ../build-tests 2>&1 | head -400
g++-13 -std=c++20 -Wall -Wextra -Werror \
-I . -I Core -I external \
Core/*.cpp external/catch2/catch_amalgamated.cpp \
Tests/TestMain.cpp Tests/CoreModelsTests.cpp Tests/LLMTranslationCoreTestsA.cpp \
Tests/TranslationPromptConfigTests.cpp Tests/VisualLineGroupingTests.cpp \
-o ../build-tests
../build-tests

Comment on lines +231 to +234
$hashes = @{
x64 = 'b56a9101f706f9d95f815f5b7fa6efbac972e86573d378b96a07cff5540c5961'
arm64 = '9a7f6d69ea6cf042ea8680b7cd0bfaa9c04f0f6cc89055d43f7f6cd0250508d3'
}

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · medium
VCLibs 的 SHA-256 哈希是脚本内联字面量,而错误信息引用 $env:VCLIBS_VERSION(env 中只定义了版本号、未定义哈希);哈希与版本两处需要同步更新却只有版本进了 env,升级 VCLibs 时容易只改 env 忘改哈希导致 CI 永久失败且报错信息误导。建议把哈希也提升到 job 级 env(如 VCLIBS_HASH_X64 / VCLIBS_HASH_ARM64),或抽到受控的清单/脚本文件并在 workflow 中引用。

Suggestion:

Suggested change
$hashes = @{
x64 = 'b56a9101f706f9d95f815f5b7fa6efbac972e86573d378b96a07cff5540c5961'
arm64 = '9a7f6d69ea6cf042ea8680b7cd0bfaa9c04f0f6cc89055d43f7f6cd0250508d3'
}
env:
VCLIBS_VERSION: 14.0.33321.0
VCLIBS_SHA256_X64: b56a9101f706f9d95f815f5b7fa6efbac972e86573d378b96a07cff5540c5961
VCLIBS_SHA256_ARM64: 9a7f6d69ea6cf042ea8680b7cd0bfaa9c04f0f6cc89055d43f7f6cd0250508d3

AtomicWriteResult out;
std::error_code ec;

const std::filesystem::path tmp = filePath.string() + ".tmp";

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · medium
writeEngineConfigFileAtomic 使用 filePath.string()(窄字符,Windows 下为 ANSI 代码页)构造临时文件路径,并用 std::remove(tmp.string().c_str()) 删除。当 %APPDATA% 路径包含非 ASCII 字符(如中文用户名)时,string() 转换可能失败或产生错误编码,导致临时文件创建失败或残留。std::filesystem 在 Windows 上原生支持宽字符,应直接使用 path 拼接(filePath += ".tmp" 或 filePath.wstring()),删除用 std::filesystem::remove 并传入 path 对象。

Suggestion:

Suggested change
const std::filesystem::path tmp = filePath.string() + ".tmp";
std::filesystem::path tmp = filePath;
tmp += ".tmp";

Comment on lines +401 to +405
// rename over an existing file may fail via std::filesystem's strict
// mapping; remove-then-rename matches MOVEFILE_REPLACE_EXISTING.
std::error_code removeEc;
std::filesystem::remove(filePath, removeEc);
std::filesystem::rename(tmp, filePath, ec);

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · high
writeEngineConfigFileAtomic 在 rename 失败时先 std::filesystem::remove(filePath) 再重试 rename;若删除成功但第二次 rename 失败(磁盘满、权限、句柄占用等),原有 engines.json 已被永久删除且临时文件随后也被 remove,用户数据丢失。真正的 MOVEFILE_REPLACE_EXISTING(或 POSIX rename)本身支持原子覆盖,不会出现『目标已删、新文件未落盘』的窗口。建议改为直接调用 Win32 MoveFileExW(MOVEFILE_REPLACE_EXISTING) 实现 Windows 覆盖语义,或至少在 remove 前保留备份。另外 tmp 路径固定为 filePath + ".tmp",多进程/多页面并发保存时会互相覆盖临时文件。

Suggestion:

Suggested change
// rename over an existing file may fail via std::filesystem's strict
// mapping; remove-then-rename matches MOVEFILE_REPLACE_EXISTING.
std::error_code removeEc;
std::filesystem::remove(filePath, removeEc);
std::filesystem::rename(tmp, filePath, ec);
// Windows 上用 MoveFileExW(MOVEFILE_REPLACE_EXISTING) 做原子覆盖,
// 避免先删除目标文件导致二次 rename 失败时数据丢失。
#ifdef _WIN32
if (!MoveFileExW(tmp.wstring().c_str(), filePath.wstring().c_str(),
MOVEFILE_REPLACE_EXISTING | MOVEFILE_WRITE_THROUGH)) {
out.error = "原子替换失败: " + std::to_string(GetLastError());
std::error_code rmEc;
std::filesystem::remove(tmp, rmEc);
return out;
}
#else
std::filesystem::rename(tmp, filePath, ec);
if (ec) { ... }
#endif

Comment on lines +69 to +78
case TranslationEngineType::Google:
case TranslationEngineType::Baidu:
// No free probe endpoint: orange without keys, gray with.
if (credentials_ && credentials_->has(credentialAccount(key.builtin))) {
return EngineHealth::Unverified;
}
return EngineHealth::Unconfigured;
case TranslationEngineType::Claude:
case TranslationEngineType::Gemini:
case TranslationEngineType::DeepL:

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · low
Google/Baidu 与 Claude/Gemini/DeepL 两个分支代码完全相同(有凭证→Unverified,否则→Unconfigured),可合并为一个 case 列表;OpenAI/Ollama 分支内部再嵌套 if (key.builtin == OpenAI) 的写法也较绕,可拆分为独立 case。

Suggestion:

Suggested change
case TranslationEngineType::Google:
case TranslationEngineType::Baidu:
// No free probe endpoint: orange without keys, gray with.
if (credentials_ && credentials_->has(credentialAccount(key.builtin))) {
return EngineHealth::Unverified;
}
return EngineHealth::Unconfigured;
case TranslationEngineType::Claude:
case TranslationEngineType::Gemini:
case TranslationEngineType::DeepL:
case TranslationEngineType::Google:
case TranslationEngineType::Baidu:
case TranslationEngineType::Claude:
case TranslationEngineType::Gemini:
case TranslationEngineType::DeepL:
if (credentials_ && credentials_->has(credentialAccount(key.builtin))) {
return EngineHealth::Unverified;
}
return EngineHealth::Unconfigured;

Comment on lines +114 to +120
std::lock_guard<std::mutex> lock(mutex_);
auto it = cache_.find(key);
if (it != cache_.end() && clock_() - it->second.at < ttl_) {
return it->second.health;
}
}
return refresh(key, token);

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · medium
health() 判断缓存过期后释放锁再调用 probe(可能耗时数秒),期间多线程会对同一 key 并发重复探测,无 in-flight 去重。建议在锁内标记探测中状态,或对探测请求做按 key 的串行化。

Comment on lines +101 to +104
if (c == '{' || c == '[') ++depth;
else if (c == '}' || c == ']') --depth;
++i;
if (depth == 0) {

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · low
手写截断恢复扫描器在遇到多余关闭符时 depth 可能减为负数,此后所有 '{'/'[' 会被当作新 item 重新以 depth=1 开始,畸形输入可能切分出错误项;建议在 depth 变为负值时终止扫描(视为尾部垃圾)。

Suggestion:

Suggested change
if (c == '{' || c == '[') ++depth;
else if (c == '}' || c == ']') --depth;
++i;
if (depth == 0) {
if (c == '{' || c == '[') ++depth;
else if (c == '}' || c == ']') {
if (--depth < 0) return std::nullopt; // 多余关闭符:尾部垃圾,终止
}
++i;
if (depth == 0) {

Comment on lines +204 to +209
auto readBox = [&b](const char* key) -> double {
if (!b.contains(key)) return 0.0;
const Json& v = b.at(key);
if (!finiteNumber(v)) return std::numeric_limits<double>::quiet_NaN();
return v.get<double>();
};

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · medium
readBox 在键缺失时返回 0.0 而非标记无效,导致 boundingBox 缺少 x/y 字段的项被当作原点处宽高正常的框参与打码,可能误覆盖左上角无关内容。在“VLM 输出不可信”的健壮性契约下,缺失关键字段应视为无效项丢弃:让 readBox 返回 optional(或 NaN),并统一走已有的 isfinite 丢弃分支。

Suggestion:

Suggested change
auto readBox = [&b](const char* key) -> double {
if (!b.contains(key)) return 0.0;
const Json& v = b.at(key);
if (!finiteNumber(v)) return std::numeric_limits<double>::quiet_NaN();
return v.get<double>();
};
auto readBox = [&b](const char* key) -> double {
if (!b.contains(key)) return std::numeric_limits<double>::quiet_NaN(); // 缺字段即无效,走统一丢弃分支
const Json& v = b.at(key);
if (!finiteNumber(v)) return std::numeric_limits<double>::quiet_NaN();
return v.get<double>();
};

Comment on lines +219 to +221
double confidence = 0.0; // missing confidence → 0 → dropped by the
// macOS >= 0.5 default filter
if (item.contains("confidence") && item["confidence"].is_number()) {

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · low
parseRegions 中 confidence 字段存在但为非 number 类型(如字符串 "0.9")时被静默置 0 丢弃,与 boundingBox 缺失字段的处理路径混杂,容易掩盖模型输出的格式漂移。建议至少在非 number 的 confidence 上与其他无效字段一致地直接 continue(跳过该项),避免依赖隐式默认值的副作用。

Suggestion:

Suggested change
double confidence = 0.0; // missing confidence → 0 → dropped by the
// macOS >= 0.5 default filter
if (item.contains("confidence") && item["confidence"].is_number()) {
double confidence = 0.0;
if (item.contains("confidence") && item["confidence"].is_number()) {

Comment on lines +42 to +44
std::string TranslationScheduler::hashTexts(const std::vector<std::string>& texts) {
// FNV-1a over the joined texts — collision-safe enough as a cache key
// discriminant alongside language + engine identity.

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · medium
resultCache_ 没有 TTL、没有大小上限,长时间运行内存无界增长;hashTexts 用无 salt 的 FNV-1a 64 位哈希,理论碰撞会导致返回错误的缓存翻译结果。建议加上容量上限(如 LRU)或在若干次运行后调用方定期 clearResultCache()。

Comment on lines +139 to +148
if (request.texts.empty()) {
SchedulerOutcome out;
TranslationResultBundle bundle;
bundle.primaryEngine =
chain.empty() ? TranslationEngineType::MtranServer : chain.front().resolvedEngine();
bundle.selectionMode = EngineSelectionMode::PrimaryWithFallback;
bundle.scene = static_cast<int>(request.scene);
out.bundle = bundle;
return out;
}

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · high
texts 为空时返回一个带 bundle 的"成功"结果,与上方注释 "empty means nothing to try, reported as an aggregated failure" 自相矛盾。调用方可能把空结果误判为翻译成功,导致空翻译结果被写入历史或 UI。建议返回聚合失败(或显式错误态)而不是空 bundle。

Suggestion:

Suggested change
if (request.texts.empty()) {
SchedulerOutcome out;
TranslationResultBundle bundle;
bundle.primaryEngine =
chain.empty() ? TranslationEngineType::MtranServer : chain.front().resolvedEngine();
bundle.selectionMode = EngineSelectionMode::PrimaryWithFallback;
bundle.scene = static_cast<int>(request.scene);
out.bundle = bundle;
return out;
}
if (request.texts.empty()) {
SchedulerOutcome out;
out.failures.push_back({TranslationEngineType::MtranServer, "", "no texts to translate"});
return out;
}

Comment on lines +157 to +158
out.failures.push_back({item.resolvedEngine(), "", "cancelled"});
return out;

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · medium
取消时把当前尚未执行的条目也记入 failures("cancelled"),与 per-engine 失败语义混淆,调用方可能把 cancelled 当作真实引擎失败展示。建议引入独立的中止状态或在失败条目上区分 cancelled。

Comment on lines +90 to +94
const std::string raw = j.at("id").get<std::string>();
c.id = fromRaw(TranslationEngineType{}, raw);
if (c.id == TranslationEngineType::Invalid) {
throw std::invalid_argument("unknown engine id: " + raw);
}

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · medium
fromJson 对缺失 "id" 直接 j.at("id") 抛异常,未知 engine id 也抛 std::invalid_argument。与同分支 EngineConfigStore.cpp『容错回退、不因单个字段失败』的解析策略不一致;engines.json 与 mac 共享,数组中单个损坏条目(人工编辑或 mac 端新增枚举)会导致整份配置解析失败。建议单条目损坏时跳过该条目而非抛出。

Suggestion:

Suggested change
const std::string raw = j.at("id").get<std::string>();
c.id = fromRaw(TranslationEngineType{}, raw);
if (c.id == TranslationEngineType::Invalid) {
throw std::invalid_argument("unknown engine id: " + raw);
}
const auto idIt = j.find("id");
if (idIt == j.end() || !idIt->is_string()) { return c; } // 跳过损坏条目
c.id = fromRaw(TranslationEngineType{}, idIt->get<std::string>());
if (c.id == TranslationEngineType::Invalid) { return c; }

Comment on lines +96 to +97
c.options = getOptional<EngineOptions>(j, "options");
c.customName = getOptional<std::string>(j, "customName");

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · low
fromJson 对缺失 "options" 字段时 getOptional(j, "options") 的行为依赖 Json 辅助实现;若 options 字段存在但类型不为对象(如写成了字符串),嵌套的 EngineOptions::fromJson 内部 j.find 等访问是否安全需要确认,避免对畸形配置抛异常。同理 toJson 中 options/customName 为空时写出 Json()(null),from_json 反序列化 null 值时也应验证不会抛出。

Suggestion:

Suggested change
c.options = getOptional<EngineOptions>(j, "options");
c.customName = getOptional<std::string>(j, "customName");
if (auto it = j.find("options"); it != j.end() && it->is_object()) {
c.options = EngineOptions::fromJson(*it);
}

Comment on lines +23 to +27
static constexpr const char* kDefaultPrompt =
"Translate the following text from {source_language} to {target_language}.\n"
"Provide only the translation without any explanations or additional text.\n\n"
"Text to translate:\n"
"{text}";

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · low
默认提示词模板在本文件(kDefaultPrompt/kDefaultInsertPrompt constexpr)与 CoreModels.cpp 的 TranslationPromptConfig::defaultPrompt()/defaultInsertPrompt() 静态字符串两处重复定义(CompatibleChatProvider.cpp、LLMChatProvider.cpp 还有第三、四处内联拼接)。目前内容一致但无单一来源,未来任何一处修改都会造成静默分叉、两端提示词不一致。建议收敛到一个共享常量(如 CoreModels 中的 static 函数,或独立 PromptTemplates.h)。

Suggestion:

Suggested change
static constexpr const char* kDefaultPrompt =
"Translate the following text from {source_language} to {target_language}.\n"
"Provide only the translation without any explanations or additional text.\n\n"
"Text to translate:\n"
"{text}";
// 引用唯一来源,避免多处拷贝导致分叉:
static const std::string& defaultPrompt();
static const std::string& defaultInsertPrompt();

Comment on lines +99 to +101
if (auto it = j.find("enginePrompts"); it != j.end() && it->is_array()) {
for (size_t i = 0; i + 1 < it->size(); i += 2) {
if (!(*it)[i].is_string() || !(*it)[i + 1].is_string()) { continue; }

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · low
TranslationPromptConfig::fromJson 中 fromRaw 失败返回 Invalid 后被跳过,这个容错是对的;但结合本文件 toJson 的写法:enginePrompts 以扁平的 [key1, value1, key2, value2, ...] 数组序列化(模拟 Swift enum-keyed dictionary),如果 value 本身是字符串这没问题;但若 JSON 中数组元素个数为奇数(外部工具手工编辑导致),循环条件 i + 1 < it->size() 会静默丢弃最后一项,行为可接受但没有任何日志/错误提示,排查线上配置问题时难以定位。建议至少在 debug 断言或注释中说明该容错语义。

Suggestion:

Suggested change
if (auto it = j.find("enginePrompts"); it != j.end() && it->is_array()) {
for (size_t i = 0; i + 1 < it->size(); i += 2) {
if (!(*it)[i].is_string() || !(*it)[i + 1].is_string()) { continue; }
if (auto it = j.find("enginePrompts"); it != j.end() && it->is_array()) {
// Flat key/value pairs; odd trailing entry or non-string pairs are
// silently skipped (tolerant decode, mirrors EngineConfigStore).
for (size_t i = 0; i + 1 < it->size(); i += 2) {
if (!(*it)[i].is_string() || !(*it)[i + 1].is_string()) { continue; }

Comment on lines +116 to +120
std::unique_ptr<DisplayMonitorService> makeDisplayMonitorService() {
static auto enumerator = makeWin32DisplayEnumerator();
static auto clock = makeSteadyClock();
return std::make_unique<DisplayMonitorService>(*enumerator, *clock);
}

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · low
工厂返回的 Service 通过裸引用持有函数级 static 的 enumerator/clock/lister/registrar,生存期依赖 exit-time 析构顺序恰好安全,属于另一处隐性生存期契约;若 Service 生命周期变化(静态单例/跨 exit 复用)将悬空。建议依赖对象改用 shared_ptr 由 Service 持有。

Suggestion:

Suggested change
std::unique_ptr<DisplayMonitorService> makeDisplayMonitorService() {
static auto enumerator = makeWin32DisplayEnumerator();
static auto clock = makeSteadyClock();
return std::make_unique<DisplayMonitorService>(*enumerator, *clock);
}
std::unique_ptr<DisplayMonitorService> makeDisplayMonitorService() {
static auto enumerator = makeWin32DisplayEnumerator();
static auto clock = makeSteadyClock();
auto service = std::make_unique<DisplayMonitorService>(*enumerator, *clock);
// Document/own the dependency lifetime explicitly rather than relying on
// exit-time destruction order.
return service;
}

…ows below

The selection overlay created a WS_EX_LAYERED window but never initialized
the layered surface (SetLayeredWindowAttributes / UpdateLayeredWindow).
An uninitialized layered window is mouse click-through everywhere - hit
testing treats the never-set alpha surface as fully transparent - so the
DComp dim kept rendering while real clicks and drags passed straight to
the windows underneath (WindowFromPoint returned Chrome). The color
picker already initializes alpha 255; do the same here.

Verified live: WindowFromPoint now resolves to the overlay at 4/4 probe
points, the dim still renders, and an injected press-move-release
sequence parks the session on the in-place toolbar (Confirmed).

Co-Authored-By: Claude Code <noreply@anthropic.com>
Comment on lines +38 to +39
- name: Add MSBuild to PATH
uses: microsoft/setup-msbuild@v2

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

security · medium
microsoft/setup-msbuild@v2 是第三方 action,仅按可变 tag 引用;tag 可被移动,存在供应链劫持风险(checkout/upload-artifact 属第一方 actions/*,按惯例可接受)。建议将 setup-msbuild 固定到完整 commit SHA。

Suggestion:

Suggested change
- name: Add MSBuild to PATH
uses: microsoft/setup-msbuild@v2
- name: Add MSBuild to PATH
uses: microsoft/setup-msbuild@<full-commit-sha>

Comment on lines +150 to +151
- name: Install compiler
run: sudo apt-get update && sudo apt-get install -y g++-13 cmake

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

performance · low
portable-check 与各 MSBuild/CMake job 每次运行都重新 apt-get install g++-13、NuGet restore 和完整 CMake 配置,未配置任何 actions/cache(NuGet 包缓存、CMake ccache、apt 缓存),拖慢 CI 并在外部源不可用时增加脆弱性。建议为 NuGet(~/.nuget/packages)和 apt 包添加缓存步骤。

Suggestion:

Suggested change
- name: Install compiler
run: sudo apt-get update && sudo apt-get install -y g++-13 cmake
- name: Cache apt packages
uses: actions/cache@v4
with:
path: /var/cache/apt/archives
key: apt-g++13-cmake
- name: Install compiler
run: sudo apt-get update && sudo apt-get install -y g++-13 cmake

Comment on lines +158 to +164
g++-13 -std=c++20 -Wall -Wextra -Werror \
-I . -I Core -I external \
Core/*.cpp external/catch2/catch_amalgamated.cpp \
Tests/TestMain.cpp Tests/CoreModelsTests.cpp Tests/LLMTranslationCoreTestsA.cpp \
Tests/TranslationPromptConfigTests.cpp Tests/VisualLineGroupingTests.cpp \
-o ../build-tests 2>&1 | head -400
../build-tests

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · low
编译错误输出经 | head -400 截断:真正的首个编译错误可能位于 400 行之后而被掩盖;若首个错误恰在末尾,则诊断信息完全丢失(只显示警告/中间噪声)。建议先输出到日志文件再 tail/grep 'error',或至少改为 tail 而非 head。

Suggestion:

Suggested change
g++-13 -std=c++20 -Wall -Wextra -Werror \
-I . -I Core -I external \
Core/*.cpp external/catch2/catch_amalgamated.cpp \
Tests/TestMain.cpp Tests/CoreModelsTests.cpp Tests/LLMTranslationCoreTestsA.cpp \
Tests/TranslationPromptConfigTests.cpp Tests/VisualLineGroupingTests.cpp \
-o ../build-tests 2>&1 | head -400
../build-tests
g++-13 -std=c++20 -Wall -Wextra -Werror ... -o ../build-tests 2> compile.log || { grep -m 20 -E 'error' compile.log; tail -50 compile.log; exit 1; }
../build-tests

Comment thread .gitignore
Comment on lines +115 to +116
# scratch repro binaries
*.obj

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · low
全局忽略 *.obj 会波及仓库任意位置(包括未来第三方源码/示例项目中的合法 .obj 文件),注释虽写明是 scratch 复现产物,但已有更精确的 /smoke_*.obj 规则覆盖该场景,二者冗余且前者过宽。建议移除裸 *.obj,保留已锚定根目录的 smoke 规则。

Suggestion:

Suggested change
# scratch repro binaries
*.obj
# scratch repro binaries (repo-root anchored)
/smoke_*.obj

Comment on lines +364 to +367
/// Columns currently laid out, derived from the measured grid width.
private var columnCount: Int {
max(1, Int(gridWidth / (cellSize + 12)))
}

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · low
columnCount 由实测 gridWidth 推导为 Int(gridWidth / (cellSize + 12)),但实际布局是 GridItem(.adaptive(minimum: 64, maximum: 96), spacing: 12):当 gridWidth 允许列宽超过 64pt 时(如 560pt 宽度下 adaptive 会排出 6 列,而此处算出 7 列),估算列数与真实列数不一致,导致上下方向键移动焦点时跳错行。建议按与布局相同的自适应规则推导列数,或改用固定列数布局使二者一致。

Comment on lines +218 to +222
for shell in ('pwsh.exe', 'powershell.exe'):
if shutil.which(shell):
result = subprocess.run([shell, '-NoProfile', '-NonInteractive', '-Command', script],
capture_output=True, text=True, check=True, timeout=60)
break

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · medium
注释声明“当 pwsh 所在环境的 Certificate PSDrive 失败时回退 powershell.exe”,但实现是 for-else + check=True:只有 shell 不在 PATH 时才会前进到下一个候选;pwsh.exe 存在但执行失败(非零退出)会直接抛 CalledProcessError 终止,不会回退到 powershell.exe,与注释描述的回退语义不符。trust_test_cert_in_root 中存在同样的模式。建议捕获 subprocess.CalledProcessError 后继续尝试下一个 shell。

Suggestion:

Suggested change
for shell in ('pwsh.exe', 'powershell.exe'):
if shutil.which(shell):
result = subprocess.run([shell, '-NoProfile', '-NonInteractive', '-Command', script],
capture_output=True, text=True, check=True, timeout=60)
break
result = None
for shell in ('pwsh.exe', 'powershell.exe'):
if not shutil.which(shell):
continue
proc = subprocess.run([shell, '-NoProfile', '-NonInteractive', '-Command', script],
capture_output=True, text=True, timeout=60)
if proc.returncode == 0:
result = proc
break
if result is None:
raise RuntimeError('certificate creation failed in every available shell')

Comment on lines +254 to +259
for shell in ('pwsh.exe', 'powershell.exe'):
if shutil.which(shell):
subprocess.run([shell, '-NoProfile', '-NonInteractive', '-Command', script],
check=True, timeout=90)
return
raise RuntimeError('Neither pwsh.exe nor powershell.exe is available')

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

security · high
trust_test_cert_in_root 会无条件把一次性自签名代码签名证书写入 LocalMachine/Root(提权环境)或 CurrentUser/Root 信任存储,且脚本没有任何清理/撤销步骤(全文件无 Remove/CertUtil -delstore 等删除逻辑)。该证书在 1 天有效期内可签署任意代码并被系统完全信任;每次运行都会生成并信任一张新证书,累积污染根信任库。建议:签名完成后在 finally 中从 Root 存储移除该证书,或改为信任到 TrustedPeople(足以满足 Add-AppxPackage 侧载校验),并文档化手动清理方式。

Suggestion:

Suggested change
for shell in ('pwsh.exe', 'powershell.exe'):
if shutil.which(shell):
subprocess.run([shell, '-NoProfile', '-NonInteractive', '-Command', script],
check=True, timeout=90)
return
raise RuntimeError('Neither pwsh.exe nor powershell.exe is available')
try:
subprocess.run(['signtool.exe', 'sign', '/sha1', thumbprint,
'/fd', 'SHA256', str(package)], check=True, timeout=60)
finally:
untrust_test_cert_in_root(thumbprint) # remove from Root store after signing

Comment on lines +1387 to +1389
files = [p for root in roots for p in root.rglob('*.[ch]pp')
if 'build' not in p.parts and 'external' not in p.parts
and 'vendor' not in p.parts]

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · low
--check-code 的文件收集使用 rglob('.[ch]pp'),该模式只匹配 .cpp/.hpp,不会匹配纯 .h 头文件。若字符串 key 只在头文件中通过 loc::t(...) 使用(例如内联函数或模板定义在 .h 中),这里会漏检,check-code 声称的"验证所有代码 key"承诺失效。建议改为 rglob('') 后按后缀集合过滤,或显式包含 *.h。

Suggestion:

Suggested change
files = [p for root in roots for p in root.rglob('*.[ch]pp')
if 'build' not in p.parts and 'external' not in p.parts
and 'vendor' not in p.parts]
files = [p for root in roots for p in root.rglob('*')
if p.suffix in ('.h', '.hpp', '.cpp', '.cc')
and 'build' not in p.parts and 'external' not in p.parts
and 'vendor' not in p.parts]

"settings.section.prompts": "Prompt Configuration",
"settings.section.shortcuts": "Keyboard Shortcuts",
"settings.shortcut.fullscreen": "Full Screen Capture",
"settings.shortcut.pick.color": "Screen Color Pick",

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · medium
该目录是 migrate_localizable.py 的 canonical 输出,运行时实际加载的是镜像副本 windows/App/Assets/Strings(AppMain.cpp 通过 loadDirectory("Assets\Strings") 读取)。两份 283 行 ×10 语言的副本若单独手改会静默漂移且无 CI 校验兜底。建议在构建/CI 中加入生成脚本的 --check 或哈希比对,确保两目录保持同步。

"settings.section.prompts": "Prompt Configuration",
"settings.section.shortcuts": "Keyboard Shortcuts",
"settings.shortcut.fullscreen": "Full Screen Capture",
"settings.shortcut.pick.color": "Screen Color Pick",

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

style · low
key 命名风格不统一:同组 shortcut key 中混用全小写点分(settings.shortcut.fullscreen、settings.shortcut.translate.and.insert)与该 key 的独立单词小写(pick.color),另外 engine.compatible.displayName、settings.vlm.baseURL、win.engines.policySaved 等使用驼峰。键集合本身在各语言间一致、可正常运行,但建议后续统一为一种风格,降低查找与引用失配风险。

Host the selection swap chain in a WS_EX_NOREDIRECTIONBITMAP window so
DirectComposition controls per-pixel alpha without an opaque layered
redirection surface. The selection remains a rectangular mouse target.

Upload the actual captured BGRA pixels when compositing preview exports,
using a separate WIC source and explicit PBGRA conversion. The old path
read the newly created empty destination instead of the source image.

Add a GUI regression driver for real mouse drags, transparent selection
content, and clipboard pixels across two sessions. Detect a locked WTS
session before driving the desktop. Add exporter pixel regressions for
annotations, padded rows, alpha conversion, and PNG round trips.

Validation: Release x64 solution build, Core 7351 assertions / 234 cases,
NativeShell 542 assertions / 106 cases, exporter 11 checks passed.
Live GUI regression is pending desktop unlock.

Co-Authored-By: Claude Code <noreply@anthropic.com>
Comment on lines +35 to +39
steps:
- uses: actions/checkout@v4

- name: Add MSBuild to PATH
uses: microsoft/setup-msbuild@v2

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

performance · medium
多个 job(msbuild /restore 全解决方案、三个独立 CMake 配置构建)每次 CI 都从零执行,且三套 CMake 构建目录(windows/NativeShell/build、windows/build、windows/build-c)互不复用,无任何 actions/cache。windows-latest 上全量构建 + 测试在 30 分钟超时内有超时风险,也造成明显资源浪费。建议为 NuGet(actions/setup-java 类似的 actions/cache 缓存 ~/.nuget/packages)和 pip 缓存,或至少合并重复的 CMake 配置。

Suggestion:

Suggested change
steps:
- uses: actions/checkout@v4
- name: Add MSBuild to PATH
uses: microsoft/setup-msbuild@v2
steps:
- uses: actions/checkout@v4
- name: Cache NuGet packages
uses: actions/cache@v4
with:
path: ~/.nuget/packages
key: nuget-${{ runner.os }}-${{ hashFiles('windows/**/*.csproj', 'windows/**/project.json') }}
- name: Add MSBuild to PATH
uses: microsoft/setup-msbuild@v2

Comment on lines +38 to +39
- name: Add MSBuild to PATH
uses: microsoft/setup-msbuild@v2

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

security · medium
第三方 action microsoft/setup-msbuild@v2 未固定到完整 commit SHA。tag 是可变的,一旦上游被劫持(或维护者误移动 tag),本工作流会在拥有 checkout 代码上下文的 runner 上执行任意代码。建议改为 microsoft/setup-msbuild@<full-sha> 并在注释中标注版本。actions/checkout 属第一方可接受,但 setup-msbuild 应固定。

Suggestion:

Suggested change
- name: Add MSBuild to PATH
uses: microsoft/setup-msbuild@v2
- name: Add MSBuild to PATH
# v2.0.0 (示例,请替换为实际 SHA)
uses: microsoft/setup-msbuild@<full-commit-sha>

Comment on lines +158 to +163
g++-13 -std=c++20 -Wall -Wextra -Werror \
-I . -I Core -I external \
Core/*.cpp external/catch2/catch_amalgamated.cpp \
Tests/TestMain.cpp Tests/CoreModelsTests.cpp Tests/LLMTranslationCoreTestsA.cpp \
Tests/TranslationPromptConfigTests.cpp Tests/VisualLineGroupingTests.cpp \
-o ../build-tests 2>&1 | head -400

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · medium
编译输出通过 | head -400 截断,在 set -euo pipefail 下有双重问题:(1) head 读满 400 行后退出,g++ 继续写 stdout/stderr 会收到 SIGPIPE,pipefail 会将非零管道状态视为步骤失败——即使编译实际成功也会偶发误报;(2) 真正的编译错误若出现在 400 行之后会被吞掉。建议移除管道(依赖 msbuild/cmake 失败时的日志裁剪可改用 2>&1 | tail -n 400 配合管道前先落盘,或直接全量输出)。

Suggestion:

Suggested change
g++-13 -std=c++20 -Wall -Wextra -Werror \
-I . -I Core -I external \
Core/*.cpp external/catch2/catch_amalgamated.cpp \
Tests/TestMain.cpp Tests/CoreModelsTests.cpp Tests/LLMTranslationCoreTestsA.cpp \
Tests/TranslationPromptConfigTests.cpp Tests/VisualLineGroupingTests.cpp \
-o ../build-tests 2>&1 | head -400
g++-13 -std=c++20 -Wall -Wextra -Werror \
-I . -I Core -I external \
Core/*.cpp external/catch2/catch_amalgamated.cpp \
Tests/TestMain.cpp Tests/CoreModelsTests.cpp Tests/LLMTranslationCoreTestsA.cpp \
Tests/TranslationPromptConfigTests.cpp Tests/VisualLineGroupingTests.cpp \
-o build-tests

Comment on lines +163 to +164
-o ../build-tests 2>&1 | head -400
../build-tests

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

performance · low
输出二进制写到 ../build-tests(仓库根目录之外),依赖 runner 工作目录布局假设;仓库根目录的 .gitignore 也未覆盖此类产物(/smoke_*.exe 等规则不匹配 build-tests)。建议输出到 windows/build/ 等仓库内已忽略目录,保持本地与 CI 行为一致。

Suggestion:

Suggested change
-o ../build-tests 2>&1 | head -400
../build-tests
-o build-tests
./build-tests

Comment thread .gitignore
Comment on lines +104 to +107
# Windows installer output and Python packaging tools
windows/dist/
__pycache__/
hubert/

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · low
hubert/ 是一条无任何注释说明的忽略规则,来源不明(可能是本地开发者目录误提交到全局 .gitignore)。全局 .gitignore 只应包含与项目构建产物相关的通用规则,个人本地目录应放进 .git/info/exclude 或全局用户配置,否则会静默忽略其他开发者可能需要的同名目录。建议删除或加注释说明用途。

Suggestion:

Suggested change
# Windows installer output and Python packaging tools
windows/dist/
__pycache__/
hubert/
# Windows installer output and Python packaging tools
windows/dist/
__pycache__/

Comment on lines +52 to +55
try {
winrt::init_apartment(winrt::apartment_type::multi_threaded);
} catch (...) {
}

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · medium
winrt::init_apartment(multi_threaded) 在线程池线程上成功初始化后没有对应的 CoUninitialize 恢复,线程复用时会累积 apartment 泄漏;同时 catch(...) 静默吞掉所有初始化异常,违反异常处理完整性原则。建议使用 RAII 包装(构造时 init_apartment,析构时 CoUninitialize),并至少记录失败的异常。

Suggestion:

Suggested change
try {
winrt::init_apartment(winrt::apartment_type::multi_threaded);
} catch (...) {
}
struct ApartmentScope {
ApartmentScope() { winrt::init_apartment(winrt::apartment_type::multi_threaded); }
~ApartmentScope() { ::CoUninitialize(); }
} apartment;

Comment on lines +87 to +88
const auto launched = ::ShellExecuteW(nullptr, L"open", uri.c_str(), nullptr, nullptr, SW_SHOWNORMAL);
if (reinterpret_cast<std::intptr_t>(launched) <= 32) {

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

bug · low
ShellExecuteW 返回值大于 32 表示成功的约定用 reinterpret_caststd::intptr_t(launched) <= 32 表达,可读性差;且 HINSTANCE 返回值强转为 intptr_t 属于实现细节依赖。建议封装具名辅助函数并加注释说明约定,提升可维护性。

Suggestion:

Suggested change
const auto launched = ::ShellExecuteW(nullptr, L"open", uri.c_str(), nullptr, nullptr, SW_SHOWNORMAL);
if (reinterpret_cast<std::intptr_t>(launched) <= 32) {
const auto launched = ::ShellExecuteW(nullptr, L"open", uri.c_str(), nullptr, nullptr, SW_SHOWNORMAL);
// ShellExecuteW: 返回值 <= 32 表示失败
if (reinterpret_cast<std::intptr_t>(launched) <= 32) {

Comment on lines +48 to +49
const std::string sceneRaw = j.at("scene").get<std::string>();
b.scene = fromRaw(TranslationScene{}, sceneRaw);

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · medium
fromJson 中 j.at("scene")/j.at("primaryEngine") 在键缺失时抛 nlohmann json::out_of_range、类型不匹配时抛 json::type_error;与文件中 defaultFor/错误路径使用的 std::invalid_argument 语义不一致,调用方若只捕获 std::invalid_argument 会漏掉这两类异常(nlohmann 的异常派生自 std::runtime_error/exception,而非 invalid_argument)。迁移损坏的历史设置文件时容易产生意外的异常传播。建议统一转换为 std::invalid_argument,或补充测试覆盖 fromJson 的缺失键/坏类型路径(当前测试只覆盖 round-trip 与 null optional)。

Suggestion:

Suggested change
const std::string sceneRaw = j.at("scene").get<std::string>();
b.scene = fromRaw(TranslationScene{}, sceneRaw);
const auto sceneIt = j.find("scene");
if (sceneIt == j.end() || !sceneIt->is_string()) {
throw std::invalid_argument("SceneEngineBinding missing string field 'scene'");
}
b.scene = fromRaw(TranslationScene{}, sceneIt->get<std::string>());

"prompt.scene.title": "Scene Prompts",
"save.success.message": "Saved to %@",
"save.success.title": "Saved Successfully",
"settings.format": "Default Format",

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maintainability · low
键前缀风格不统一:settings.format/settings.language/settings.* 与 win.settings.* 并存,history.copy.translation 与 win.tray.history 等也分属不同前缀体系。键集已验证跨 10 个语言文件完全一致(无拼写问题),但前缀混用会使后续新增键的归属不明确。建议在迁移脚本(migrate_localizable.py)或文档中约定统一前缀规则,避免长期键集漂移。

Suggestion:

Suggested change
"settings.format": "Default Format",
"settings.format": "Default Format", // 建议统一为 win.settings.* 前缀或在新键中遵循单一前缀约定

"win.shortcuts.title": "Global shortcuts",
"win.tray.checkUpdates": "Check for updates",
"win.tray.closeAll": "Close all",
"win.tray.colorHistory": "Color history",

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

style · low
键名命名风格不一致:"win.tray.colorHistory" 使用驼峰式(colorHistory),而同组其它键均为全小写点分风格(如 "win.colorhistory.*"、"win.tray.pickColor" 除外一致)。两个前缀 "win.colorhistory" 与 "win.tray.colorHistory" 还分别指代同一功能的窗口标题与托盘菜单项,后续维护易混淆或拼错。建议统一为小写点分(如 win.tray.colorhistory),并在 migrate_localizable.py 的键映射表中同步修正。

Suggestion:

Suggested change
"win.tray.colorHistory": "Color history",
"win.tray.colorhistory": "Color history",

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant